Energy companies' carbon reduction, low-carbon transformation, and green innovation for deep learning algorithms under the carbon neutrality goal

To address the problems faced by energy companies, such as complex carbon emission data modeling and insufficient decision support, as well as the difficulty of traditional methods in meeting both real-time and accuracy requirements, this paper proposes a carbon emission modeling framework based on a graph neural network. This paper first constructs a heterogeneous graph with production equipment, energy consumption units, and emission factors as nodes, and defines edge weights through material/energy flows; then, this paper uses multi-scale graph convolution to aggregate equipment layer and plant layer features, introduces a dynamic attention mechanism, calculates node weights, and combines emission thresholds to enhance key node identification, embeds a gated recurrent unit to model temporal evolution, and fuses graph features and historical states through reset gates and update gates; finally, this paper establishes a dual-objective optimization function for total carbon emissions and production costs, balancing the two with weight coefficients. Experiments show that the F1 value of the equipment warning performance of the proposed model reaches 0.89; under the 30% emission reduction constraint, the marginal cost is 209.41 yuan/ton, which is 115.38 yuan/ton lower than that of the graph convolutional networks model. This cost advantage lowers the threshold for green innovation for enterprises; by coordinating plant-level energy dependencies, this method provides a technical path for coordinated emission reduction in upstream and downstream industrial chains, and provides accurate modeling and optimized decision-making support for the low-carbon transformation of energy enterprises.